P1022 Network performance

[TerryBarnaby]

We are using a P1022 linked with a Virtex6 FPGA on one of our products. The P1022 is working well.

However the network performance with respect to CPU usage is a bit of an issue. We are using the Linux 3.3.2 kernel. Testing TCP/IP performance using iperf it uses 100% of one of the CPU's cores just to send data at 821 Mbits/sec and, separately, receive data at 612 Mbits/sec. The Gigabit Ethernet link's data rate is thus limited by the processor performance and there is nothing left to do other tasks. We wonder if this is due to the Linux Ethernet driver efficiency issues or if it is just down to the P1022's raw performance.

Any ideas/info ?

[lunminliang]

How many cores and how many flows do you use when testing? What is your frame length? As frame length would be an important factor to the line rate you get especially when less than 512bits.

There is no networking benchmark for P1022 available, but there is some performance data for P1020 which is similar to P1022. On P1020 one core one flow, the IPV4 forwarding performance is 216Mbps for 64B frame size, and 988Mbps for 1518B frame size. These data is based on Linux P1020RDB 2.6.32.

[TerryBarnaby]

This is with two cores enabled but really only one core running as there is just one iperf process running, nothing else of note. The MTU is set to 1500 and iperf fills the packets (has a buffer size of 8k) so frame length about 1500 bytes and data length a little less. With an MTU of 9000 data rate versus CPU usage is obviously much better.

Your figures for the P1020 seem to match what we are seeing.

So the question remains, is this high CPU usage a Linux Ethernet driver/stack/interrupts efficiency issue (We could look at optimising the code) or is it just down to the P1022's raw performance. An Intel 2GHz Xeon uses about 6% CPU on one core doing the same (not like for like, I know). Would the P20xx series be significantly better than the P1022 at this ?

Any ideas/info ?

[lunminliang]

I do not think it's down to the P1022's raw performance, as the benchmark performance shows, it's my personal view. The P2020 is similar at this.

[TerryBarnaby]

So do you think it is primarily Linux Ethernet driver/TCPIP stack software issue rather than the hardware ?

We are considering looking at the Linux driver performance but was hoping for indications of this would be worthwhile before commencing on any work.